docs: record parallel fork budget and the mktemp-uniqueness constraint by Chemaclass · Pull Request #814 · TypedDevs/bashunit

Chemaclass · 2026-07-16T03:27:46Z

🤔 Background

Follow-up to #812/#813: the fork-budget rule predates the parallel-mode profiling and says "per-test cost is fork-free", which is inaccurate for --parallel (one mktemp per test remains, for a provable reason).

💡 Changes

Adds the perf(runner): stop forking per test when publishing parallel results #813 replacement patterns (parent pre-creates the workers' dir; skip sanitizing on empty input) and the measured parallel budget (61 → 21 forks per 10-test run).
Documents why the per-test mktemp can't become a deterministic name on Bash 3 ($$/RANDOM shared by subshells, BASHPID is 4.0+) and two explicit don't-fix notes (wait_for_job_slot, spinner).

Updates the fork-budget rule with the #813 results: the parallel per-test publishing patterns (parent pre-creates the workers' dir; skip sanitizing on empty input), the measured parallel budget (61 -> 21 forks per 10-test run), why the remaining per-test mktemp cannot become a deterministic name on Bash 3 (subshells share $$ and RANDOM state; BASHPID is 4.0+), and two explicit don't-fix notes (wait_for_job_slot, spinner).

Chemaclass · 2026-07-16T03:27:46Z

🤔 Background

Follow-up to #812/#813: the fork-budget rule predates the parallel-mode profiling and says "per-test cost is fork-free", which is inaccurate for --parallel (one mktemp per test remains, for a provable reason).

💡 Changes

Adds the perf(runner): stop forking per test when publishing parallel results #813 replacement patterns (parent pre-creates the workers' dir; skip sanitizing on empty input) and the measured parallel budget (61 → 21 forks per 10-test run).
Documents why the per-test mktemp can't become a deterministic name on Bash 3 ($$/RANDOM shared by subshells, BASHPID is 4.0+) and two explicit don't-fix notes (wait_for_job_slot, spinner).

Updates the fork-budget rule with the #813 results: the parallel per-test publishing patterns (parent pre-creates the workers' dir; skip sanitizing on empty input), the measured parallel budget (61 -> 21 forks per 10-test run), why the remaining per-test mktemp cannot become a deterministic name on Bash 3 (subshells share $$ and RANDOM state; BASHPID is 4.0+), and two explicit don't-fix notes (wait_for_job_slot, spinner).

Chemaclass · 2026-07-16T03:27:46Z

🤔 Background

Follow-up to #812/#813: the fork-budget rule predates the parallel-mode profiling and says "per-test cost is fork-free", which is inaccurate for --parallel (one mktemp per test remains, for a provable reason).

💡 Changes

Adds the perf(runner): stop forking per test when publishing parallel results #813 replacement patterns (parent pre-creates the workers' dir; skip sanitizing on empty input) and the measured parallel budget (61 → 21 forks per 10-test run).
Documents why the per-test mktemp can't become a deterministic name on Bash 3 ($$/RANDOM shared by subshells, BASHPID is 4.0+) and two explicit don't-fix notes (wait_for_job_slot, spinner).

Updates the fork-budget rule with the #813 results: the parallel per-test publishing patterns (parent pre-creates the workers' dir; skip sanitizing on empty input), the measured parallel budget (61 -> 21 forks per 10-test run), why the remaining per-test mktemp cannot become a deterministic name on Bash 3 (subshells share $$ and RANDOM state; BASHPID is 4.0+), and two explicit don't-fix notes (wait_for_job_slot, spinner).

Chemaclass · 2026-07-16T03:27:46Z

🤔 Background

Follow-up to #812/#813: the fork-budget rule predates the parallel-mode profiling and says "per-test cost is fork-free", which is inaccurate for --parallel (one mktemp per test remains, for a provable reason).

💡 Changes

Adds the perf(runner): stop forking per test when publishing parallel results #813 replacement patterns (parent pre-creates the workers' dir; skip sanitizing on empty input) and the measured parallel budget (61 → 21 forks per 10-test run).
Documents why the per-test mktemp can't become a deterministic name on Bash 3 ($$/RANDOM shared by subshells, BASHPID is 4.0+) and two explicit don't-fix notes (wait_for_job_slot, spinner).

Updates the fork-budget rule with the #813 results: the parallel per-test publishing patterns (parent pre-creates the workers' dir; skip sanitizing on empty input), the measured parallel budget (61 -> 21 forks per 10-test run), why the remaining per-test mktemp cannot become a deterministic name on Bash 3 (subshells share $$ and RANDOM state; BASHPID is 4.0+), and two explicit don't-fix notes (wait_for_job_slot, spinner).

Chemaclass · 2026-07-16T03:27:46Z

🤔 Background

Follow-up to #812/#813: the fork-budget rule predates the parallel-mode profiling and says "per-test cost is fork-free", which is inaccurate for --parallel (one mktemp per test remains, for a provable reason).

💡 Changes

Adds the perf(runner): stop forking per test when publishing parallel results #813 replacement patterns (parent pre-creates the workers' dir; skip sanitizing on empty input) and the measured parallel budget (61 → 21 forks per 10-test run).
Documents why the per-test mktemp can't become a deterministic name on Bash 3 ($$/RANDOM shared by subshells, BASHPID is 4.0+) and two explicit don't-fix notes (wait_for_job_slot, spinner).

Updates the fork-budget rule with the #813 results: the parallel per-test publishing patterns (parent pre-creates the workers' dir; skip sanitizing on empty input), the measured parallel budget (61 -> 21 forks per 10-test run), why the remaining per-test mktemp cannot become a deterministic name on Bash 3 (subshells share $$ and RANDOM state; BASHPID is 4.0+), and two explicit don't-fix notes (wait_for_job_slot, spinner).

Chemaclass · 2026-07-16T03:27:46Z

🤔 Background

Follow-up to #812/#813: the fork-budget rule predates the parallel-mode profiling and says "per-test cost is fork-free", which is inaccurate for --parallel (one mktemp per test remains, for a provable reason).

💡 Changes

Adds the perf(runner): stop forking per test when publishing parallel results #813 replacement patterns (parent pre-creates the workers' dir; skip sanitizing on empty input) and the measured parallel budget (61 → 21 forks per 10-test run).
Documents why the per-test mktemp can't become a deterministic name on Bash 3 ($$/RANDOM shared by subshells, BASHPID is 4.0+) and two explicit don't-fix notes (wait_for_job_slot, spinner).

Updates the fork-budget rule with the #813 results: the parallel per-test publishing patterns (parent pre-creates the workers' dir; skip sanitizing on empty input), the measured parallel budget (61 -> 21 forks per 10-test run), why the remaining per-test mktemp cannot become a deterministic name on Bash 3 (subshells share $$ and RANDOM state; BASHPID is 4.0+), and two explicit don't-fix notes (wait_for_job_slot, spinner).

Chemaclass · 2026-07-16T03:27:46Z

🤔 Background

Follow-up to #812/#813: the fork-budget rule predates the parallel-mode profiling and says "per-test cost is fork-free", which is inaccurate for --parallel (one mktemp per test remains, for a provable reason).

💡 Changes

Adds the perf(runner): stop forking per test when publishing parallel results #813 replacement patterns (parent pre-creates the workers' dir; skip sanitizing on empty input) and the measured parallel budget (61 → 21 forks per 10-test run).
Documents why the per-test mktemp can't become a deterministic name on Bash 3 ($$/RANDOM shared by subshells, BASHPID is 4.0+) and two explicit don't-fix notes (wait_for_job_slot, spinner).

Updates the fork-budget rule with the #813 results: the parallel per-test publishing patterns (parent pre-creates the workers' dir; skip sanitizing on empty input), the measured parallel budget (61 -> 21 forks per 10-test run), why the remaining per-test mktemp cannot become a deterministic name on Bash 3 (subshells share $$ and RANDOM state; BASHPID is 4.0+), and two explicit don't-fix notes (wait_for_job_slot, spinner).

Chemaclass · 2026-07-16T03:27:46Z

🤔 Background

Follow-up to #812/#813: the fork-budget rule predates the parallel-mode profiling and says "per-test cost is fork-free", which is inaccurate for --parallel (one mktemp per test remains, for a provable reason).

💡 Changes

Adds the perf(runner): stop forking per test when publishing parallel results #813 replacement patterns (parent pre-creates the workers' dir; skip sanitizing on empty input) and the measured parallel budget (61 → 21 forks per 10-test run).
Documents why the per-test mktemp can't become a deterministic name on Bash 3 ($$/RANDOM shared by subshells, BASHPID is 4.0+) and two explicit don't-fix notes (wait_for_job_slot, spinner).

Updates the fork-budget rule with the #813 results: the parallel per-test publishing patterns (parent pre-creates the workers' dir; skip sanitizing on empty input), the measured parallel budget (61 -> 21 forks per 10-test run), why the remaining per-test mktemp cannot become a deterministic name on Bash 3 (subshells share $$ and RANDOM state; BASHPID is 4.0+), and two explicit don't-fix notes (wait_for_job_slot, spinner).